[Epic][P0] Refocus AgentMesh as the framework-neutral Agent Control Plane

[0YHR0]

Product decision

AgentMesh must stop competing as another multi-Agent framework. Its durable differentiation is the
control plane between stochastic Agents and deterministic real-world systems:

Bring any Agent. Run it reliably. Govern every action. Observe and recover everything.

LangGraph, OpenAI Agents SDK, CrewAI, Codex, Claude Code, A2A peers, and custom workers should
become managed runtimes—not competing architecture centers. Virtual Company, Office, Music Studio,
and Market Intelligence remain optional experiences built on the control plane.

Problem

The repository already contains strong reliability and governance primitives, but the public story
and composition boundary are diluted across orchestration, vertical applications, game UI, runtime,
and platform operations. A new user cannot tell what must be adopted, what is optional, or why
LangGraph/LangSmith alone is insufficient.

Target architecture

• Control plane: Agent identity/version, Task/Run/Attempt, scheduling, policy, Permit, budget,
  audit, and recovery.
• Runtime plane: framework-neutral adapters for local, isolated, and remote A2A Agents.
• Capability plane: governed MCP tools and declarative capability/Skill bundles.
• Experience plane: optional UI and business scenarios installed as extensions.

Workstreams

1. Framework-neutral Managed Agent Runtime contract and conformance suite.
2. A non-LangGraph adapter proving “bring your own Agent”.
3. Governed Action/Permit protocol and SDK.
4. Reliability model plus reproducible failure/chaos benchmark.
5. Agent Principal and delegated identity chain.
6. Isolated runtime execution profiles.
7. Fleet Operations Console for 20–100+ Agents.
8. Move vertical experiences out of the core composition boundary.
9. Governed capability/Skill bundle contract.
10. Relate rollout proposal [Proposal] Add first-class rollout groups for isolated, concurrent, verifiable agent execution #26 to isolated/verifiable execution instead of creating a parallel
    runtime model.

Delivery backlog

P0 — prove the control plane

• [P0] Define framework-neutral Managed Agent Runtime API v0.1 #135 — Managed Agent Runtime API v0.1
• [P0] Prove framework neutrality with a non-LangGraph managed Agent #136 — non-LangGraph managed Agent proof
• [P0] Specify Governed Action and scoped Permit Protocol v0.1 #137 — Governed Action and scoped Permit Protocol v0.1
• [P0] Publish the AgentMesh Reliability Model and chaos benchmark #138 — Reliability Model and chaos benchmark

P1 — harden and productize

• [P1] Introduce Agent Principal and delegated workload identity #139 — Agent Principal and delegated workload identity
• [P1] Add provider-neutral isolated Agent runtime execution #140 — provider-neutral isolated Agent runtime execution
• [P1] Build a Fleet Operations Console for 20–100+ Agents #141 — Fleet Operations Console for 20–100+ Agents
• [P1] Extract flagship experiences from the public control-plane core #142 — extract flagship experiences from public core
• [P1] Define governed Capability Bundle (Skill) contract v0.1 #143 — governed Capability Bundle (Skill) contract v0.1

Related existing proposal

• [Proposal] Add first-class rollout groups for isolated, concurrent, verifiable agent execution #26 — rollout groups; consume [P0] Define framework-neutral Managed Agent Runtime API v0.1 #135/[P1] Add provider-neutral isolated Agent runtime execution #140 isolation and [P0] Publish the AgentMesh Reliability Model and chaos benchmark #138 benchmark semantics

Acceptance criteria

• README first screen and architecture map identify AgentMesh as an Agent Control Plane within ten
  seconds.
• A runtime other than LangGraph passes the same lifecycle/governance conformance suite.
• Every irreversible tool action can be explained as Intent → Policy → Approval (optional) →
  Permit → Execution → Reconciliation.
• A published benchmark demonstrates convergence after duplicate delivery, worker death, lease
  expiry, Redis interruption, and unknown external-write outcome.
• Core starts and passes tests without importing vertical experience runtime/UI packages.
• Console answers which Agents exist, which versions/runtimes they use, what they are doing,
  permissions, spend, failures, approvals, and revocation state.
• Minimal single-Agent operation remains simple and requires no optional experience.

Non-goals

• Building a better planner prompt, ReAct loop, or model-specific harness.
• Replacing LangGraph, MCP, A2A, Kubernetes, or sandbox providers.
• Making the Office/game UI the core product.
• Claiming production readiness without reproducible evidence.
• Adding new vertical scenarios before the control-plane boundary is proven.

[0YHR0]

Design baseline is now in draft PR #144: ADR 0007, Managed Runtime API v0.1, Governed Action Protocol v0.1, reliability/chaos qualification, and the Luna implementation plan. This PR intentionally closes no implementation issues.